基于低秩模型的联合特征学习与识别算法研究

Low-rank modeling generally refers to a class of methods that solves problems by representing variables of interest as low-rank matrices. It has been a research hotspot in machine learning and pattern recognition, which has achieved great success in various fields such as biometrics, image understanding and video analysis. However, most of existing low-rank modeling algorithms performed as pre-processing methods to do denoising or graph construction and thus cannot well match the subsequent feature extraction and recognition tasks. Motivated by such scientific problem, four main themes are put forward in the research proposal based on the joint learning framework: 1) research on the joint low-rank modeling and subspace learning algorithm towards the subspace learning in noisy environment; 2) research on the joint low-rank learning and subspace clustering algorithm instead of the existing manner which treats two tasks separately; 3) research on the joint low-rank learning and semi-supervised learning algorithm by using the label information to enhance the discriminative power of the graph constructed by low-rank models; and 4) research on the joint low-rank feature learning and classification methods to let the feature learning process better match the classification process. The significance of this research proposal will not only build some joint feature learning and recognition algorithms for image understanding and biometrics, but also enrich the contents in low-rank modeling family. Expected outcomes of this research proposal will provide novel techniques and fundamental theoretical basis for more effectively processing the high dimensional data; furthermore, it will induce a new theoretical perspective for investigating joint learning and recognition algorithms in machine learning and pattern recognition field.

近年来，低秩学习模型在生物特征识别、图像处理与视频分析等应用中取得了显著的效果，已经成为机器学习与模式识别领域的研究热点。然而目前的低秩学习算法多数被用作数据去噪预处理或图的构造方法等，低秩学习目标不能很好的与数据的后续特征学习、识别等任务协同起来。因此，本项目拟在联合学习的框架下开展以下内容研究：（1）针对噪声环境下的子空间学习问题，提出联合低秩学习与子空间学习模型；（2）针对噪声环境下的图构造与子空间聚类问题，提出联合低秩学习与子空间聚类模型；（3）针对引入部分标签信息以增强低秩系数图的判别性问题，提出联合低秩学习与半监督学习模型；（4）针对数据特征学习和分类任务独立分阶段进行的问题，提出联合低秩特征学习与分类模型。项目旨在通过以上内容的研究，提出一套完整的联合特征学习与识别算法框架，为实现更有效地高维数据分析提供理论和算法支撑，为噪声环境下的协同特征学习与识别算法研究开拓新的视角。

低秩学习模型在生物特征识别、图像处理与视频分析等应用中取得了显著的效果，已经成为机器学习与模式识别领域的研究热点。然而目前的低秩学习算法多数被用作数据去噪预处理或图的构造方法等，低秩学习目标不能很好的与数据的后续特征学习、识别等任务协同起来。因此，在联合学习的框架下实现基于低秩学习的联合特征提取与识别，对实现更有效地高维数据分析提供理论与算法支撑，具有重要的科学意义。.本项目对基于低秩模型的联合特征学习与识别算法展开了多方面的研究。为避免先得到低秩系数图再进行图的子空间学习两阶段模式的次优性，提出了基于低秩表示与谱回归的联合子空间鲁棒学习算法，实现了图的构造与子空间投影矩阵的协同学习、噪声数据的自动恢复以及模型的可扩展性；为刻画数据可能存在的非线性信息以及在核空间中进一步发掘数据的结构化信息，提出了流形自适应核低秩表示模型；为进一步地提升图上的相关机器学习算法性能，提出了联合秩约束的结构化图学习与无监督特征选择算法，联合秩约束的结构化图学习与图正则化的非负矩阵分解算法，联合秩约束的结构化图学习与自表示概念分解算法，解决了在给定图的情况下，各学习算法的性能严重依赖给定图的质量的问题；此外，项目组还提出了稀疏保持的超限学习机算法、正交超限学习机算法、并行向量场约束的非负矩阵分解算法等。.项目组在执行过程中，对低秩学习相关基础理论、算法、技术进行研究并获得了较丰富的成果。目前已经发表论文16篇，其中SCI期刊论文7篇，CCF推荐B类会议论文6篇，普通国际会议论文2篇，中文核心期刊论文1篇；申请专利3项，授权1项；参与获得中国电子学会技术发明三等奖1项。部分成果发表在国内外重要期刊与学术会议如Complexity、Neurocomputing、Computer Methods and Programs in Biomedicine、Journal of Ambient Intelligence and Humanized Computing、ICASSP、BIBM等。此外，部分研究成果还处于审稿中。接下来，将围绕低秩学习在情感脑电数据的有效成分发掘与识别、声呐图像的目标检测进行应用拓展研究。